Detection of standing dead trees using leaf-on and leaf-off UAV-borne laser scanning point cloud data in mixed forests

Krašovec, Nina (author), Repe, Blaž (mentor), Höfle, Bernhard (comentor)

The assessment of forest health is gaining importance with the increasing frequency and severity of drought events. Forest ecosystems are becoming more vulnerable and susceptible to diseases and insect attacks, leading to increased tree mortality and risk of fire. Light Detection and Ranging (LiDAR) allows to obtain valuable 3-dimensional geometrical and spectral information, which is becoming more widely used in forest health assessments. The aim of this study was to examine the potential of bi-temporal UAV-borne laser scanning (ULS) data and voxel-based metrics for predicting the occurrence of standing dead trees without tree delineation. The position of each standing dead tree was measured during field inventory. ULS data was collected under leaf-on and leaf-off conditions. A 2D moving window approach was developed to extract feature sets based on cells, height bins, and columns from leaf-on and leaf-off datasets. The classification was carried out using a random forest classifier, resulting in accuracies of 0.87 and 0.86 for the two tested plots. These results were achieved using the bi-temporal dataset and using a threshold of 2.5 m for the distance of the 2D window position and the location of the field measurement. Among the two voxel-based groups of metrics (height bins and columns), height bins provided more valuable information from vertical vegetation strata, which improved the classification of live and dead trees. Metrics derived from the point cloud divided into columns proved to be prone to noisy points and missing data. The presented approach using a 2D moving window gave satisfactory results, but further analysis would be needed in different forest settings to determine the effects of vegetation structure on classification performance.

• Language: English
• Keywords: LiDAR, drone-based, tree mortality, bi-temporal data, random forest
• Work type: Master's thesis/paper
• Organization: FF - Faculty of Arts
• Year: 2021
• PID: 20.500.12556/RUL-128432
• Publication date in RUL: 13.07.2021

Secondary language

• Language: Slovenian
• Title: Zaznavanje odmrlih stoječih dreves z uporabo laserskega skeniranja v obdobjih rasti in mirovanja

Abstract

Ocenjevanje zdravstvenega stanja gozdov postaja vse pomembnejše zaradi pogostejših in daljših sušnih obdobij. Gozdni ekosistemi postajajo bolj ranljivi in dovzetnejši za bolezni in napade žuželk, zaradi česar se povečujeta smrtnost dreves in nevarnost požarov. Z metodo LiDAR (ang. Light Detection and Ranging) lahko pridobimo trirazsežne geometrijske in spektralne podatke, ki se vse pogosteje uporabljajo pri ocenjevanju zdravstvenega stanja gozdov. Namen te študije je bil proučiti potencial bitemporalnih podatkov laserskega skeniranja z uporabo brezpilotnega zrakoplova (ang. UAV-borne laser scanning ali ULS) in atributov na podlagi vokslov za napovedovanje pojavljanja odmrlih stoječih dreves brez njihove predhodne segmentacije. Lokacijo vsakega odmrlega stoječega drevesa smo izmerili na terenu. Podatke laserskega skeniranja smo zajeli v obdobjih rasti in mirovanja. Za pridobivanje atributov na podlagi celic, višinskih slojev in stolpcev iz bitemporalnih podatkov smo uporabili drsno okno. Klasifikacijo smo izvedli z uporabo algoritma naključnih gozdov, pri čemer smo dosegli klasifikacijsko točnost 0,87 za proučevano območje P1 in klasifikacijsko točnost 0,86 za proučevano območje P2. To smo dosegli z bitemporalnimi podatki in z uporabo praga 2,5 m za razdaljo med položajem drsnega okna in najbližjo lokacijo terenske meritve. Med obema skupinama atirbutov, ki temeljita na vokslih (višinski sloji in stolpci) smo z višinskimi sloji pridobili bolj uporabne podatke iz navpičnih slojev vegetacije, kar je izboljšalo klasifikacijo živih in odmrlih dreves. Izkazalo se je, da so atributi, pridobljeni iz oblaka točk razdeljenega v stolpce, dovzetni za šum in manjkajoče vrednosti. Z uporabo drsnega okna smo dobili zadovoljive rezultate, vendar bi bile za ugotavljanje vpliva vegetacijske sestave na uspešnost klasifikacije potrebne nadaljnje analize v različnih gozdnih okoljih.

• Keywords: LiDAR, brezpilotni zrakoplov, umrljivost dreves, bitemporalni podatki, naključni gozd